Skip to content

Fix profiling in Python 3.10#259

Merged
jamadden merged 3 commits intomasterfrom
issue256
Sep 1, 2021
Merged

Fix profiling in Python 3.10#259
jamadden merged 3 commits intomasterfrom
issue256

Conversation

@jamadden
Copy link
Contributor

This required some additional state management and several new tests.

Fixes #256.

The cframe member was always being initialized from the PyThreadState's root_cframe, which doesn't properly have use_tracing set, so new greenlets didn't get tracing.

Here, we do exactly what CPython's ceval.c does and stack-allocate a new CFrame object in g_initialstub and copy the use_tracing value from the currently executing thread state.

This seems to solve the tracing issue, but further tests are needed; none of the existing tests failed with the broken behaviour. In addition to adding an example from #256, we need to test that the use_tracing flag properly propagates up and down when a profiler is set inside a greenlet, likewise I've only been testing switch, need to test throw, and need to test several variations on calling g.run(): as a method, just a target function, an object assigned to a greenlet, etc.
This proves that propagating the tracer *upward* (to a parent greenlet) in Cpython 3.10 fails. However, 4287a40 did fix propagating it downward (to a new greenlet). Likely, propagating among sibling greenlets is also broken.
Because it's stack based, we need to store it separately.
@jamadden jamadden merged commit d70ab45 into master Sep 1, 2021
@jamadden jamadden deleted the issue256 branch September 1, 2021 13:25
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Change in profiler behaviour in CPython 3.10

1 participant